Amazon Bedrock AgentCore powers protein research copilot for peptide search

Amazon has demonstrated a new protein research copilot built on Amazon Bedrock AgentCore and the Strands Agents SDK that tackles the slow, error-prone process of finding structurally similar peptides across large datasets. The system transforms natural language queries like "Find 10 similar peptides to LPAIVREAI" into structured searches by using a dedicated LLM-as-parser pattern. It then generates 960-dimensional protein embeddings via an ESM-C 300M model (deployed as a SageMaker AI serverless endpoint) and performs cosine similarity search against peptide embeddings stored in Amazon Aurora PostgreSQL with pgvector. A second LLM agent summarizes results with scientific context.

The architecture runs on a single Bedrock AgentCore runtime orchestrating three specialized tools: the parser, the searcher, and the summarizer. A Streamlit frontend on AWS Fargate provides the conversational interface. The solution deploys in 30–45 minutes and demonstrates how to combine serverless ML inference, vector databases, and multi-agent orchestration for domain-specific research assistants. Using Anthropic Claude Sonnet 4.6 via Bedrock Converse API, the copilot can filter by species and limit results, producing downloadable tables and actionable insights from the IEDB virus epitope dataset.